Cosegmentation Loss: Enhancing segmentation with a Few Training Samples by Transferring Region Knowledge to Unlabeled Images

Wataru Shimoda, Keiji Yanai

Feb 17, 2017 (modified: Feb 21, 2017) ICLR 2017 workshop submission readers: everyone
  • Abstract: We treat semantic segmentation where a few pixel-wise labeled samples and a large number of unlabeled samples are available. For this situation we propose cosegmentation loss which enables us to transfer the knowledge of a few pixel-wise labeled samples to a large number of unlabeled images. In the experiments, we used human-part segmentation with a few pixel-wise labeled images and 1715 unlabeled images, and proved that the proposed co-segmentation loss helped make effective use of unlabeled images.
  • TL;DR: Co-Segmentation Loss for semi-supervised semantic segmentation
  • Conflicts: